Skip to content

UCSD Qualcomm Institute

2026.01~02

FinBERT vs. LLM sentiment classification research

Published at KCC 2026Second author of six

Pseudo-label sentiment: does the threshold keep the source distribution? The lower the confidence threshold α, the closer it stays to the original financial distribution Negative Neutral Positive Source (financial) Financial PhraseBank test, 4,846 12.5 59.4 28.1 α = 0.75 · 7 cycles Nearly matches the source 12 58 30 α = 0.9 · 3 cycles Negatives triple; drifts from the source 31 47 22 At α = 0.75: Macro F1 0.85 / 0.84 → 0.87, Macro Accuracy 0.86 → 0.88 Values in %. Rounded, so a row may not sum to 100. Source: unpublished follow-up, not the KCC 2026 paper. Evaluating the Model Stability of FinBERT through SSL under Cross-Domain Mismatch, Team C

Problem

It was not known what happens when FinBERT, trained only on a financial corpus (FPB), is applied as-is to general reviews.

My part

Wrote and ran the BERT and FinBERT baselines. Also owned the data pipeline that filtered 650K Yelp reviews with a Gemini quality filter.

Result

Showed negative transfer with numbers and published it at KCC 2026 as second author of six. A follow-up with semi-supervised self-training improved scores on the financial test set.
PythonPyTorchHugging FaceFinBERTGemini APIOpenAI APIpandas
Paper